1
Throughput Optimization for Massive MIMO
Systems Powered by Wireless Energy Transfer
Gang Yang, Chin Keong Ho, Rui Zhang, and Yong Liang Guan
Abstract—This paper studies a wireless-energy-transfer
(WET) enabled massive multiple-input-multiple-output
(MIMO) system (MM) consisting of a hybrid data-and-
arXiv:1403.3991v2 [cs.IT] 2 Jan 2015
energy access point (H-AP) and multiple single-antenna
users. In the WET-MM system, the H-AP is equipped
with a large number M of antennas and functions like
a conventional AP in receiving data from users, but
additionally supplies wireless power to the users. We
consider frame-based transmissions. Each frame is divided
into three phases: the uplink channel estimation (CE) phase,
the downlink WET phase, as well as the uplink wireless
information transmission (WIT) phase. Firstly, users use a
fraction of the previously harvested energy to send pilots,
while the H-AP estimates the uplink channels and obtains
the downlink channels by exploiting channel reciprocity.
Next, the H-AP utilizes the channel estimates just obtained
to transfer wireless energy to all users in the downlink via
energy beamforming. Finally, the users use a portion of the
harvested energy to send data to the H-AP simultaneously
in the uplink (reserving some harvested energy for sending
pilots in the next frame). To optimize the throughput and
ensure rate fairness, we consider the problem of maximizing
the minimum rate among all users. In the large-M regime,
we obtain the asymptotically optimal solutions and some
interesting insights for the optimal design of WET-MM
system.
Index Terms—Massive MIMO, wireless energy transfer,
energy beamforming, channel estimation, throughput max-
imization, asymptotic analysis
I. I NTRODUCTION
Recently, far-field wireless energy transfer (WET) has
emerged as a promising technology to address energy and
lifetime bottlenecks for power-limited devices in wireless
networks [1] [2]. WET refers to using the radiative elec-
tromagnetic (EM) wave emitted from a power transmitter
to deliver energy to a power receiver (see [1] [2] and
references therein). Since EM waves decay quickly over
distances, to realize WET in practice, the EM energy
needs to be concentrated into a narrow beam to achieve
efficient transmission of power, also referred to as energy
beamforming [3].
G. Yang and Y. L. Guan are with the School of Electrical and
Electronic Engineering, Nanyang Technological University, Singapore (e-
mail:{yang0305, eylguan}@ntu.edu.sg).
C. K. Ho is with the Institute for Infocomm Research, A⋆ STAR,
Singapore (e-mail: hock@i2r.a-star.edu.sg).
R. Zhang is with the Department of Electrical and Computer Engineer-
ing, National University of Singapore (e-mail: elezhang@nus.edu.sg). He
is also with the Institute for Infocomm Research, A⋆ STAR, Singapore.
Simultaneous wireless information and power transfer
(SWIPT) that was proposed in [3] [4], has been extensively
studied in literature, since it offers great convenience to
mobile users with concurrent data and energy supplies.
The authors in [5], [3] studied the performance limits
of single-input-single-output (SISO) and multiple-input-
multiple-output (MIMO) SWIPT systems, respectively, and
characterized various achievable rate-energy (R-E) trade-
offs by practical receiver designs. SWIPT has also been
studied in fading channels [6], orthogonal frequency divi-
sion multiplexing (OFDM) systems [7]–[10], and multiuser
channel setups such as broadcast channels [11], [12], relay
channels [13], [14], and interference channels [15], [16].
Moreover, [17] studied a hybrid network which overlays
an uplink cellular network with randomly deployed power
beacons that charge users wirelessly, while [18] studied a
similar setup in cognitive radio networks with secondary
users harvesting wireless power opportunistically from
nearby primary users’ transmission.
Another emerging trend focuses on the study of us-
ing wireless power to support wireless communications,
thus forming a wireless powered communication network
(WPCN) [19]. In a WPCN, an access point (AP) with
multiple antennas first transfers energy to multiple single-
antenna users via downlink beamforming, then the users
use the harvested energy to perform uplink wireless infor-
mation transmission (WIT) to the AP. Due to the channel
propagation loss, the harvested energy decays exponen-
tially with respect to the distance between users and the AP.
Hence, the challenge for deploying a WPCN in practice
is to ensure that the uplink WIT is feasible when users
are solely powered by downlink WET. Some commercial
products on WPCN are currently available [20]. Moreover,
several technological advancements provide impetuses to
the mass deployment of WPCN in the near future. First,
the ongoing deployment of small-cell networks like fem-
tocell [21] shortens the energy transfer distance, and thus
decreases the channel propagation attenuation significantly.
Second, the advancement in low-power electronics like the
ultra-low-power transceiver [22] is substantially reducing
the required uplink transmission power. Third, the use of
concurrent downlink WET and uplink WIT was investi-
gated for multiple users in [23]. Finally, massive antenna
arrays [24] [25] with tens to hundreds of elements can be
used to form sharp energy beams towards users to achieve

a point-to-point WET efficiency of close to one. Massive
MIMO has been prototyped and shown to enormously
improve the transmission capacity by exploiting its large
array gain at the base stations (BSs) [24]. However, little
is exactly known on its ability of enhancing the efficiency
and distance of WET, which is studied in this work.
In a WPCN, the total time duration is divided into the
downlink WET and the uplink WIT. Assuming perfect
CSI, the single-user scenario was studied in [26], [27].
With finite-rate feedback, [26] optimized the time duration
for downlink WET to maximize a lower bound on the
uplink WIT rate. [27] maximized the energy efficiency of
uplink WIT, by jointly optimizing the time duration and
transmit power for downlink WET. Also assuming perfect
CSI, the multiuser WPCN was studied in [19], [28]. In
particular, [19] considered the scenario in which users use
the harvested energy to send independent information to
the AP through time division multiple access (TDMA). The
sum throughput was maximized subject to user fairness by
jointly optimizing the time allocation for downlink WET
and uplink WIT. [28] maximized the minimum throughput
between all users and the AP equipped with multiple
antennas, by optimizing the downlink energy beamformer,
the uplink transmit power, and searching the optimal time
allocation for downlink WET.
In practice, perfect CSI at the transmitter is not available
due to various factors such as channel estimation error,
feedback error, and time-varying channel. The knowledge
of accurate CSI is however especially important in a
WPCN. More accurate CSI contributes to both higher
efficiency of energy transfer and higher uplink information
rate. Typically, the AP needs to perform channel estimation
(CE) first. Although a longer time duration for CE leads
to more accurate CSI available at the AP, it reduces the
WET and WIT duration, which may lead to less harvested
energy and lower throughput. To optimize the throughput,
there is thus a design freedom, that is the time spent for
CE, WET and WIT.
For a point-to-point WET system, the effect of CE
and feedback on energy beamforming was studied in our
previous work [29] and [30]. In particular, [29] investigated
the dynamic allocation of time resource for CE and energy
resource for WET. The optimal preamble length is obtained
by solving a dynamic programming problem. The solution
is a threshold-type policy that depends only on the channel
estimate power, and hence allows a low-complexity WET
system to be implemented in practice. [30] studied transmit
energy beamforming by using one-bit feedback from the
energy receiver, to facilitate hardware implementation.
Based on the one-bit feedback information, the energy
transmitter adjusts transmit beamforming and concurrently
obtains improved estimates of the channel to the energy
receiver.
In this paper, we consider a WET-enabled massive
2
MIMO system (termed WET-MM), which consists of one
hybrid data-and-energy access point (H-AP) with constant
power supply and equipped with a large-scale antenna
array, and a set of distributed single-antenna users that
rely on the wireless power sent from the H-AP for uplink
transmission. We assume frame-based transmissions with
the time-division-duplexing (TDD) protocol. Each frame
is divided into three phases: the uplink CE (i.e., channel
estimation) phase, the downlink WET (i.e., wireless energy
transfer) phase, as well as the uplink WIT (i.e., wireless
information transmission) phase. Firstly, users use a frac-
tion of the harvested energy in the previous frames to send
pilots, while the H-AP estimates the uplink channels and
obtains the downlink CSI by exploiting channel reciprocity.
Secondly, the H-AP transfers wireless energy to all users in
the downlink via energy beamforming with appropriately
designed weights. Thirdly, the users use the rest of the
harvested energy (after reserving energy for the CE in
the next frame) to send their independent information
to the H-AP simultaneously in the uplink. To optimize
the throughput and ensure rate fairness, we consider the
problem of maximizing the minimum rate among all users.
The design variables are the time allocations for uplink
CE and downlink WET (subject to a given total time for
CE, WET and WIT of each frame), the energy allocation
weights in the downlink WET phase for different users, as
well as the fraction of energy used for CE (versus WIT)
at each user.
To the best of our knowledge, this paper is the first in the
literature to consider the WET-MM system with imperfect
CSI. To investigate the optimality of the proposed WET-
MM system with imperfect CSI, we compare it to the
ideal case with perfect CSI known at the H-AP. Here, we
introduce a metric, namely, the massive MIMO degree-of-
rate-gain (MM-DoRG), which is defined as the asymptotic
scaling order of the uplink rate with respect to log M , i.e.,
R
κ , lim , (1)
M→∞ log M
where M is the number of transmit antennas at the H-
AP, and R is the data rate that depends on M . The
definition of MM-DoRG is motivated by two observations.
First, the optimal practical system with imperfect CSI
achieves a rate that asymptotically scales by a factor of
two with respect to log M . Second, the asymptotic rate of
a massive MIMO system powered by energy broadcasting
without beamforming scales according to log M . These
two observations will be shown in Section V. Hence, the
MM-DoRG can be interpreted as the proportional gain of
the asymptotic rate for a massive MIMO system powered
by WET with beamforming, with respect to a massive
MIMO system powered by energy broadcasting without
beamforming. The terminology of MM-DoRG is similar
to the well-known degree-of-freedom (DoF). In this paper,

we use a lower bound on the achievable rate for analytical
tractability, which is numerically shown to be tight. We
focus on the use of zero-forcing (ZF) detection in the
uplink WIT phase. We show that the proposed WET-MM
system has the following advantages:
• In terms of MM-DoRG, the proposed WET-MM
system is optimal, as it achieves the same MM-DoRG
as the ideal case with perfect CSI, i.e., κWET−MM =
κIdeal = 2. For a massive MIMO system powered by
energy broadcasting without beamforming, the MM-
DoRG is one. Energy beamforming is thus shown
to be crucial for the proposed WET-MM system to
achieve high-efficiency WET and the optimality in
terms of MM-DoRG.
• The proposed WET-MM system achieves the best
possible rate fairness among users, as all users asymp-
totically achieve a common rate in the large-M
regime. This asymptotically solves the problem of rate
unfairness due to the “double near-far” effect in a
WPCN [19].
• The proposed WET-MM system is of low complexity
and low overhead, requiring only conventional beam-
forming and detection at the H-AP. Also, there is no
overhead (control signaling) required for indicating
the WET and WIT operation, since the asymptotically
optimal time and energy allocation for all users is
derived and then fixed in all frames.
• Numerical results corroborate the analysis. It is nu-
merically shown that the asymptotic analysis and the
asymptotically optimal solution, although obtained
under the large-M assumption, is accurate for M as
small as 25. Moreover, to achieve a desired common
rate for all users with a given maximal AP-user
distance, the proposed WET-MM system is numer-
ically shown to require less antennas at the H-AP,
roughly a square root of that required by a massive
MIMO system powered by energy broadcasting with-
out beamforming.
The rest of this paper is organized as follows: Section II
presents the system model. The problem formulation is
then given in Section III. We derive the achievable rate in
Section IV. The asymptotic analysis is given in Section V,
followed by the asymptotically optimal solutions given in
Section VI. Numerical results are given in Section VII.
Finally, Section VIII concludes the paper.
Notation: Scalars are denoted by letters (or Greek let-
ters), vectors by boldface lower-case letters, and matrices
by boldface upper-case letters. I and 0 denote an identity
matrix and an all-zero vector, respectively, with appropriate
dimensions. For a matrix A of arbitrary size, A∗ , AT , AH
denote the conjugate, the transpose and the conjugate
transpose of A, respectively. For a diagonal matrix D
1
of order K, D 2 denotes the diagonal matrix whose k-
th diagonal entry is the square root of the k-th diagonal
3
Frame 1 Frame 2 ... Frame ...
(UL) Channel Estimation (DL) Wireless Energy Transfer (UL) Wireless Information Transmission
(Energy for Pilot: r Ek ) (Harvested Energy by User k: Ek ) (Energy for Data: (1 - r ) Ek )
tT aT (1 - t - a )T
T
Fig. 1: Frame Structure
entry of D. E[·] denotes the statistical expectation. The
distribution of a circularly symmetric complex Gaussian
(CSCG) random vector with mean u and covariance matrix
Σ is denoted by CN (u, Σ). “∼” stands for “distributed
as”. kak2 denotes the Euclidean norm of a complex vector
a. O(·) denotes the big-O order. “−−−→” denotes the
convergence as M → ∞.
II. S YSTEM M ODEL
We consider a wireless powered communication network
(WPCN) consisting of a hybrid data-and-energy access
point (H-AP) with M antennas, and K single-antenna
users. Each user uses the harvested energy to power its
uplink information transmission. We assume that the H-AP
and all users are perfectly synchronized and operate with
a TDD protocol. We consider frame-based transmissions
over flat-fading channels on a single frequency band.
As shown in Fig. 1, the length of one frame is fixed as
T seconds, which is assumed to be less than the coherence
interval. Each frame consists of three phases. In the first
CE phase of time period τ T (0 < τ < 1) seconds,
the users send orthogonal training pilots, and the H-AP
estimates the uplink channels and obtains the downlink CSI
by exploiting channel reciprocity. Then in the second WET
phase of time period αT (0 < α < 1) seconds, the H-AP
delivers energy via beamforming, and the users harvest
energy from the received RF signals. In the final WIT
phase of the remaining time period (1 − τ − α)T seconds,
all users transmit information to the H-AP simultaneously
in a space-division-multiplexing-access (SDMA) manner.
Clearly, the constraint 0 ≤ τ + α ≤ 1 should be satisfied.
For convenience, we normalize T = 1 in the rest of this
paper without loss of generality. Let G = [g1 g2 · · · gK ]
be the uplink M ×K channel matrix between the H-AP and
the K users, i.e., [G]mk = gmk is the channel coefficient
between the m-th antenna of the H-AP and the k-th user.
We model the matrix G as
1
G = HB 2 , (2)
where H is the M × K matrix of independent Rayleigh
fading coefficients between the H-AP and K users, i.e.,
[H]mk = hmk ∼ CN (0, 1), and B is a K × K diagonal
matrix, with [B]kk = βk denoting the (long-term) path
loss of the channel between the H-AP and user k that is

assumed to be constant over frames and taken to be known
a priori at both the H-AP and user k.
A. Uplink Channel Estimation Phase
In the CE phase, all users simultaneously transmit mu-
tually orthogonal pilot sequences of length L (L ≥ K)
symbols, which allows the H-AP to estimate the channels.
In practice, L should be chosen such that the constraint
LTs ≤ τ is satisfied, in which Ts is the sampling period.
User k transmits a pilot sequence with power pCE k . Define
D = diag{LpCE CE CE
1 , Lp2 , · · · , LpK }. The pilot sequences
used by K users can be represented by an L × K matrix
1
ΦD 2 , where Φ is of size L × K and satisfies ΦH Φ = IK
to preserve orthogonality of the pilots [31]. The received
signal at the H-AP is thus given by
 T
1
Yp = G ΦD 2 + N, (3)
where N is an M × L matrix with independent and
identically distributed (i.i.d.) elements each distributed as
CN (0, σ 2 ). Given Yp , the minimum mean-square-error
b = [b
(MMSE) estimate of G, denoted by G b2 · · · g
g1 g bK ],
is given by [32]


Gb = Yp Φ∗ BD + σ 2 IK −1 D 21 B. (4)
Denote the estimation error by E , G b − G. From the
property of MMSE estimation, E is independent of G. b
From (4), the elements of the k-th column of the matrix
E are random variables with zero mean and variance
2 2 βk
σe,k = σe,k (L, pCE
k )= . (5)
1 + βk LpCE
k /σ
2
B. Downlink Energy Transfer Phase
In the downlink WET phase, the M × 1 transmitted
√ b where pdl is the transmit
signal is given by pdl w(G),
power for downlink WET, and the beamformer w(G) b is to
b
be designed depending on the channel estimate G, subject
b 2 = 1. Let the user noise n0,k ∼ CN (0, σ 2 ).
to kw(G)k 0 (1 − ρ)Ek (α, ρ, ξk ). Thus, the transmit power is given by
Assuming channel reciprocity, the received baseband signal
in one symbol period at user k is written as
√ 1−τ −α
zk = pdl gkH w(G)b + n0,k . (6)
We assume the energy due to the ambient noise in (6)
cannot be harvested. We also assume the energy harvested
by user k in one symbol period equals the energy of the
equivalent baseband signal in (6) by the law of energy
conservation; our results remain valid if a fixed energy
loss, or a fixed fraction energy loss, is incurred. Hence,
the expected harvested energy by user k is given by
 
2
Qk (L, pCE , α, w( b = αE b pdl gH w(G)
G)) b . (7)
k G,G k
4
We further obtain the asymptotically optimal energy beam-
former in the following Lemma 1.
Lemma 1. As M → ∞, the asymptotically optimal
beamformer that maximizes the harvested energy in (7) is
a linear combination of the normalized channel estimates
between the H-AP and users, i.e.,
K p
X bk
g
b =
w(G) ξk , (8)
kb
gk k 2
k=1
PK
where the weights ξk ’s are subject to k=1 ξk = 1.
Proof: See proof in Appendix A.
The weights ξk ’s represent the energy allocation for
downlink WET among users, which will be optimized later
in Section VI. In the sequel, we use Qk (L, pCE
k , α, ξk ) to
denote the harvested energy by user k, since the beam-
former in (8) depends on only the weights ξk ’s, given G. b
We assume that users have infinite battery storage for
storing the harvested energy. Note that the energy used
for pilot transmission is drawn from the harvested energy
in previous frames. At steady state, we assume a fraction
ρ (0 < ρ < 1) of the harvested energy Qk (L, pCE k , α, ξk )
is used by users to send pilots. The amount of pilot energy
affects the accuracy of CSI and thus the efficiency of
downlink WET. Under this energy-splitting scheme, the
harvested energy for user k, denoted as Ek (α, ρ, ξk ), will
be derived in Section IV.
C. Uplink Data Transmission Phase
In the uplink WIT phase, the received baseband signal
vector at the H-AP is given by
y = Gx + n, (9)
√
where x = [x1 x2 · · · xK ]T with xk = pk sk , where pk
is the transmit power of user k, and sk ∼ CN (0, 1) is its
information-carrying signal that is independent of signals
of other users. Here, the noise vector n ∼ CN (0M , σ 2 IM ).
Taking into account of the energy consumption for
uplink pilots, the energy left for uplink transmission is
(1 − ρ)Ek (α, ρ, ξk )
pk = pk (τ, α, ρ, ξk ) = . (10)
In the uplink WIT phase, all users simultaneously
transmit to the H-AP. The H-AP adopts a linear detector
A = [a1 a2 · · · aK ] to detect the information for all users.
Specifically, a ZF or MRC detector which performs well
for large M [24], is given by
  −1
 Gb G b HG b , for ZF
A= (11)
 b
G, for MRC.
From (9), the signal after using the detector A is obtain
as r = AH Gx + AH n. In particular, the detected signal
 5

associated with user k, denoted by rk , is written as OP-MM system is of low complexity. The H-AP first
K broadcasts energy in all
√ directions, i.e., the weight vector
√ X √ H
rk = pk aH b
g s + bi si
pi ak g w = [1 1 · · · 1]T / M . Then it performs uplink CE,
k k k
i=1,i6=k followed by uplink WIT. A fraction ρ of the harvested
K energy is used for CE, with the rest of the energy used for
X √ H
− pi ak ei si + aH WIT. Hence, channel estimation is employed only for data
k n, (12)
i=1 detection, but not for WET.

where ai , gbi , and ei are the i-th column of A, G b and

E, respectively. The last three terms in (12) are treated III. P ROBLEM F ORMULATION
as interference and noise, and assumed to be Gaussian
distributed to maximize entropy. Hence, the achievable rate Before formulating the problem, we first present the
of the uplink transmission from user k is given by fairness issue in a WPCN. Due to the distance-dependent
  signal attenuation, the users that are far from the H-AP
Rk = Rk (τ, α, ρ, ξk ) , (1−τ −α)E log (1+γk ) , (13) harvest less energy, and moreover, they have to perform
where the signal-to-interference-plus-noise-ratio (SINR) uplink transmission using higher power to overcome the
propagation. This is referred as the “double near-far”
pk |aH b k |2
k g effect [19], which intuitively leads to user unfairness, with
γk = .
P
K PK
2 + |aH a |σ 2 significantly lower rates achieved by far users. In our
pi |aH b
g
k i | 2 + |aH a |
k k p σ
i e,i k k
i=1,i6=k i=1 framework, the rates of users can be balanced by adjusting
the energy allocation weights ξ = [ξ1 ξ2 · · · ξK ]T of the
The exact expression for the rate Rk is not analytically
beamformer in (8).
tractable. Instead, we use an analytically tractable lower
To optimize the throughput and achieve user fairness, we
bound on the rate achieved by using the harvested energy.
maximize the minimum rate for all users, by optimizing
By utilizing the convexity of the function log (1 + 1/x)
over the CE time τ , the WET time α, the energy-splitting
and Jensen’s inequality, the rate Rk is lower-bounded (see
fraction ρ, and the energy allocation vector ξ. We have the
the same deviation as in Section III-A2 in [31]) as
following problem formulation
ek , (1 − τ − α) log (1 + γ
Rk ≥ R ek ) , (14)
(P1 ) max ek (τ, α, ρ, ξk )
min R (15a)
τ, α, ρ, ξ 1≤k≤K
where the SINR eγk is given by the expression
K
X
  −1
P Hb 2 H
PK
2 H 2
s. t. ξk = 1 (15b)
  p |a g
i k i | +|a a
k k | p σ
i e,i +|a a
k k |σ  k=1
 i6=k i=1 
E   . 0≤τ +α≤1 (15c)
  pk |aH
k b
g k | 2 
τ ≥ 0, α ≥ 0, ξk ≥ 0, ∀k (15d)
0 ≤ ρ ≤ 1, (15e)
We shall observe numerically in Section VII that the lower
bound in (14) is tight for massive MIMO systems. In the where (15b) is due to the power constraint at the H-AP
sequel, we use Rek as the achievable rate. for energy beamforming, and (15c) is due to the frame
duration constraint. The achievable rate Rek (τ, α, ρ, ξk ) is
D. Ideal Case and OP-MM System given in (14).
To verify the optimality, we shall compare the proposed For fixed M , solving Problem (P1 ) is challenging, be-
WET-MM system with the ideal case, in which the H- cause of the nonlinear relationship between the information
AP is assumed to have perfect CSI for downlink energy rate and the variables τ, α, ρ and ξ. In our system model,
beamforming and uplink information decoding. The CE the variables are coupled in the uplink CE, downlink WET
2
phase in Fig. 1 is thus removed, i.e., τ = 0, ρ = 0, σe,k = and uplink WIT phases as follows:
0, ∀k. Clearly, the WET-MM system with imperfect CSI • The CE time τ and the fraction ρ of the harvested
cannot perform better than the ideal case. energy used for pilots affect the CSI accuracy.
To investigate the performance improvement by intro- • The CSI accuracy affects the amount of harvested
ducing energy beamforming in the WET-MM system, energy in the downlink WET phase, the signal power
we also consider the massive MIMO system powered and the interference power for data detection in the
by energy broadcasting without beamforming, namely the uplink WIT phase.
omnidirectional powering massive MIMO (OP-MM) sys- • Besides the CSI accuracy, the amount of harvested
tem. The energy broadcasting is being widely used in energy also depends on the energy allocation weights
commercial RF-energy-harvesting products [20], [22]. The ξ and the WET time α.
 6

• Moreover, the interference from other users depends sides, which can be written as a quadratic equation in
on the energy harvested by them, the time allocation Qk (L, pCE CE
k , α, ξk ). Solving for Qk (L, pk , α, ξk ) and dis-
τ, α, as well as the energy-splitting fraction ρ. carding the negative solution, we obtain the harvested
Nevertheless, it is possible to obtain interesting asymp- energy for user k in (17). From (5), the variance of channel
totical solutions and insights in the large-M regime. In the estimation error is obtained as in (18).
sequel, we obtain analytical expression for the achievable
rate in Section IV, followed by the asymptotic analysis B. Achievable Rate for WET-MM System with Imperfect
in Section V, and then obtain the asymptotically optimal CSI
 −1
solutions in Section VI. b G
For ZF detection, the detector is A = G b HG
b ,
with aH
i gbj equals 1 if i = j and equals 0 otherwise. The
IV. A NALYSIS ON ACHIEVABLE R ATE
achievable rate is further obtained in the following lemma.
Before deriving the achievable rate in Section IV-B, we b the harvested
Lemma 4. With MMSE channel estimate G,
first obtain the harvested energy.
energy Ek (α, ρ, ξk ) in (17), and M ≥ K +1, an achievable
uplink rate of user k for ZF detection is given by
A. Harvested Energy
eZF = (1 − τ − α) log (1 + γ
R ek ) , (19)
Using the beamformer in (8), the harvested energy is k
derived in the following lemma. where the SINR
Lemma 2. With the beamformer in (8), the expected (M − K)βk2 ρEk (α, ρ, ξk )
harvested energy by user k is given by γ
ek = ,
  σ2

1−τ −α PK
βi Ei (α,ρ,ξi )
(M −1)σ 2 σ 2 βk ρ+ Ek (α,ρ,ξk) 1−ρ + βi ρEi (α,ρ,ξk )+σ 2
CE
Qk (L, pk , α, ξk ) = αpdl ξk βk M 1− i=1
M (βk LpCE 2
k +σ ) Proof: Following similar steps as in the proof for
+ αpdl βk (1 − ξk ). (16) Proposition 7 in [31], we have
Proof: See Appendix B. ekZF = (1 − τ − α)·
R (20)
The first term in (16) is the harvested energy from the  
beam directed toward user k, while the second term rep- 2


 pk (τ, α, ρ, ξk )(M −K) βk −σe,k (α, ρ, ξk ) 
resents the energy harvested from beams directed toward 
log 1+ .
P
K 
other users but still harvested by user k. The expected σ2 + 2
pi (τ, α, ρ, ξi )σe,i (α, ρ, ξi )
harvested energy Qk (L, pCE k , α, ξk ) increases, as either the i=1
(discrete) pilot length L or the pilot power pCE k increases. The result in (19) is thus obtained, by substituting the error
The term LpCE k in (16) represents the energy used by variance in (18) into (20).
user k for pilot transmission. By assumption, a fraction ρ of For MRC detection, from (11), we have A = G. b The
the expected harvested energy is used for pilots. With this achievable rate is obtained in Lemma 5.
energy-splitting scheme, we further obtain the harvested Lemma 5. With MMSE channel estimate G, b MRC detec-
energy and the variance of channel estimation error in tion and M ≥ 2, an achievable UL rate of user k is given
Lemma 3. by
Lemma 3. With the beamformer in (8) and using the

ReMRC = (1 − τ − α) log 1 + γ ekMRC , (21)
fraction ρ of the expected harvested energy for pilots, the k
harvested energy by user k is where the SINR
q
2
gk (α, ρ, ξk )+ gk2 (α,ρ,ξk )+ 4αpρdl σ ekMRC =
γ
Ek (α, ρ, ξk ) = , (17)
2 (M − 1)βk2 ρEk (α, ρ, ξk )
σ2

 σ2 (1−τ−α) P  ,
where the function βk ρ + Ek (α,ρ,ξ + β i Ei (α, ρ, ξi ) +β k σ 2
k ) 1−ρ
i6=k
σ2
gk (α, ρ, ξk ) = αpdl βk (ξk (M − 1) + 1) − .
βk ρ where the harvested energy Ek (α, ρ, ξk ) is given by (17).
Moreover, the variance of the channel estimation error of Proof: See Appendix C.
user k is rewritten as For ZF detection, the interference in (19) is due to only
βk σ 2 the channel estimation error, while for MRC detection,
2 2
σe,k = σe,k (α, ρ, ξk ) = . (18) the interference component in (21) comes from both the
βk ρEk (α, ρ, ξk ) + σ 2
channel estimation error and the multi-user interference
Proof: Replacing LpCE CE
k in (16) by ρQk (L, pk , α, ξk ), (MUI). Comparatively, MRC detection has lower complex-
we obtain the equation with Qk (L, pCE
k , α, ξ )
k in both ity of computation, since all antennas independently apply
 7

bk to maximize the signal power for user

a matched filter g Then, from (17), the asymptotic harvested energy at user
k. k is given by
Ek (α, ξk , ρ) −−−→ Ekasym (α, ξk ) , αpdl βk ξk M. (27)
C. Achievable Rate for Ideal Case and OP-MM System
2
1) Ideal Case: With perfect CSI, i.e., σe,k = 0, the ideal Remark 1 (Discussion on the asymptotically harvested
case is the special case for the analysis in Sections IV-A energy). The asymptotically harvested energy Ekasym is
and IV-B. With the beamformer in (8), the harvested energy achieved when M is sufficiently large such than the M -
by user k is given by dependent term in (17) is dominant over other terms. We
note that the Ekasym is independent of ρ, as the ρ-dependent
Ek (α, ξk ) = αpdl ξk βk M + αpdl βk (1 − ξk ). (22) terms in (17) are negligible.
Moreover, the achievable rate is given by For the condition (26), we assume that α, ρ and ξk
are arbitrarily fixed and independent1 of M . However, we
ek,Ideal = R
R ek,Ideal (α, ξk ) (23) noted that when α = O(M −2ν ) for 0 < ν < 12 , the
  
 E (α,ξ )(M−K)β condition (26) still holds, and Ekasym (α, ξk ) = O(M 1−2ν )
 (1−α) log 1+ ,
k k k
 (1−α)σ2 for ZF
! increases as M increases. This observation will be used in
= Section VI.
 P Ek (α,ξk )(M−1)βk
(1−α) log 1+ Ei (α,ξi )βi +(1−α)σ2 , for MRC.

Also, we observe that the asymptotically harvested en-
i6=k
ergy in (27) approaches the harvested energy in (22) for
2) OP-MM System: It can be shown that the harvested the ideal case with perfect CSI. That is, with the proposed
energy by user k is Ek,OP-MM (α) = αpdl βk . By replacing scheme of energy splitting between UL CE and UL WIT,
Ek (α, ξk , ρ) in (19) by Ek (α), the achievable rate for ZF the energy beamforming with estimated CSI asymptotically
is given by achieves the DL-WET performance limit that is achieved


ek,OP-MM
R ZF
(τ, α, ρ) = (1−τ −α) log 1 + γ ZF
ek,OP-MM , by the ideal case.
(24) Moreover, the asymptotic rate is given in Theorem 1 for
where the SINR ZF, and in Theorem 2 for MRC.

ZF (M − K)αpdl βk3 ρ Theorem 1. For fixed τ, α, ρ and K, when M is suffi-

γ
ek,OP-MM =   . ciently large such that (26) is satisfied, the asymptotically
σ2

1−τ −α P K
αpdl βi2
2
σ βk ρ + αpdl βk 1−ρ + αpdl βi2 ρ+σ2 achievable rate of user k for ZF detection is given by
i=1 !
Similarly, from (21), the achievable rate for MRC is thus eZF M (M − K)αpdl βk2 ξk ρ
Rk −−−→ (1 − τ − α) log 1 +  −α)ρ 
.
given by σ 2 K + (1−τ1−ρ

(28)
ReMRC
k,OP-MM (τ, α, ρ) = (1−τ −α) log 1 + γ eMRC ,
k,OP-MM
(25) Proof: Taking M → ∞ and assuming (26) holds,
where the SINR from (19) in Lemma 4 and (27), the asymptotically achiev-
able rate is derived in (28).
MRC
γ
ek,OP-MM =
Theorem 2. For fixed τ, α, ρ and K, when M is
(M − 1)αpdl βk3 ρ
! . sufficiently large such that
σ2

σ2 (1−τ −α) P
K ( )
βk ρ + αpdl βk 1−ρ + αpdl βi2 + βk σ 2 σ2 σ 2 (1 − τ − α)
i=1,i6=k M ≫ max max , ,
1≤k≤K αpdl ρβk2 ξk αpdl βk ξk (1 − ρ)
V. A SYMPTOTIC A NALYSIS (29)
In this section, we analyze the UL achievable rate in the the asymptotically achievable rate of user k for MRC
large-M regime. detection is given by
 
A. Asymptotic Analysis for WET-MM System with Esti- eMRC −−−→ (1−τ −α) log  (M − 1)βk2 ξk 
Rk 1 + P 2  . (30)
mated CSI βi ξi
i6=k
To obtain analytical insights, we consider the massive
MIMO regime where the number of transmit antennas at Proof: Taking M → ∞, and assuming M ≫
2
H-AP, M , is sufficiently large. Specifically, we assume max αpσdl(1−τ −α)
βk ξk (1−ρ) and (26) holds, from (27), the noise
1≤k≤K
σ2
M ≫ max 2ξ . (26) 1 Later
1≤k≤K αpdl ρβk k we shall consider α, ρ, ξk as variables to be optimized.
 8

power in (21) in Lemma 4 is negligible compared to αpdl ξk βk M . Then, the achievable rate for ZF is given
the MUI power, and the asymptotically achievable rate is from (23) by
further derived in (30).  2

For MRC detection, the asymptotic rate is independent ReZF (α, ξk ) −−−→ (1 − α) log 1 + αpdl βk ξk M (M − K) .
k
σ 2 (1 − α)
of the energy-splitting fraction ρ. This is because both the
(34)
signal power and the MUI power in (21) are proportional
to ρ. The effect of ρ is thus cancelled. Thus, ρ can be The achievable rate for MRC is given from (23) by
arbitrarily chosen in (0, 1). Also, we note that the condition  
in (29) can be easily satisfied as long as ρ does not  (M −1)β 2 ξ 
approach 1. eMRC (α, ξ) −−−→ (1−α) log 
R 1+ K k k
 . (35)
k
For the proposed WET-MM system,  P 

the asymptotic βi2 ξi
SINR in (28) for ZF is of order O M 2 , since the MUI is i=1,i6=k
cancelled. For MRC, however, the asymptotic SINR in (30)

The asymptotic SINR of ZF in (34) is of order O M 2 ,
is of order O (M ), due to the MUI. The factor (M − 1)
while the asymptotic SINR for MRC in (35) is of order
in the asymptotical SINR is due to the maximum ratio
O (M ). By definition, the maximal asymptotic MM-DoDR
combining. The MM-DoDR of the proposed WET-MM
is thus obtained in (33).
system is thus given immediately by the following theorem.
Remark 2 (Advantages over the OP-MM system). The
Theorem 3. For a WET-MM system, the maximal asymp- proposed WET-MM system outperforms the OP-MM sys-
totic MM-DoDR of any user k is given by tem in the following three aspects. First, the MM-DoRG
(
2, for ZF achieved by the WET-MM system κWET−MM is double
κWET−MM = (31) of that achieved by the OP-MM system κOP−MM . This
1, for MRC.
is because more harvested energy (due to energy beam-
Proof: Clearly, from Theorem 1 and Theorem 2, we forming) leads to both more accurate CSI and increased
have uplink transmit power in the WIT phase. From Theorem 3
( and Theorem 4, we can interpret the MM-DoRG metric
ek
R 2(1 − τ − α), for ZF
κWET-MM = lim = as the proportional gain of the asymptotic rate for a
M→∞ log M 1 − τ − α, for MRC. massive MIMO system power by WET with beamforming,

with respect to the OP-MM system. Second, we shall
For ZF, let τ → 0, α = O M −2ν → 0, where 0 < ν <
1 −ϕ see in Remark 3 in Section VI that the proposed WET-
2 . For MRC, let τ → 0, and α = O (M ) → 0, where
MM system asymptotically achieves the best possible rate
0 < ϕ < 1. The maximal asymptotic MM-DoRG is thus
fairness among users (i.e., a common rate), while no
obtained in (31).
fairness is guaranteed in the OP-MM system. Third, as
will be numerically shown in Section VII, to achieve a
B. Asymptotic Analysis for Ideal Case and OP-MM System desired common rate for all users with a given maximal
The MM-DoRG for the ideal case and the OP-MM AP-user distance, the proposed WET-MM system requires
system is given in the following theorem. less antennas at the H-AP, roughly a square root of that
required by the OP-MM system.
Theorem 4. The maximal asymptotic MM-DoDR for the
OP-MM system is given by
VI. A SYMPTOTICALLY O PTIMAL S OLUTION
κOP−MM = 1, for ZF and MRC. (32)
Following the asymptotic analysis in Section V, in this
The maximal asymptotic MM-DoDR for the ideal case of section, we derive the asymptotically optimal solution to
the WET-MM system is given by the minimum rate maximization Problem (P1 ) in Sec-
( tion VI-A for the proposed WET-MM system, and in
2, for ZF
κIdeal = (33) Section VI-B for the ideal case and the OP-MM system.
1, for MRC.

Proof: For the OP-MM system, the asymptotic achiev-
able rate is still given in (24) for ZF, and in (25) for
MRC. Note that the asymptotic SINR is of order O (M ).
By definition, the maximal asymptotic MM-DoDR for the
OP-MM system is obtained in (32).
For the ideal case of the WET-MM system, from (22),
we obtain that the harvested energy Ek (α, ξk ) −−−→
A. Asymptotically Optimal Solution for WET-MM Systems
1) Asymptotically Optimal Energy Allocation Weights:
In the large-M regime, the optimal ξk that maximizes the
minimum rate is given in Lemma 6.
Lemma 6. When M satisfies the condition in (26), for
both ZF and MRC detection, the minimum achievable rate
 9

is maximized when the energy allocation weight is chosen Lemma 8. For ZF detection, when M is sufficiently large
as to satisfy (26), the asymptotically optimal time allocation
1 for CE and WET is given by
βk2
ξk⋆ = . (36)

P
K
1
⋆
τZF −−−→ 0, α⋆ZF = O M −2ν −−−→ 0,
βi2 where ν > 0, and ν → 0. (38)
i=1

Proof: See Appendix D. Proof: Suppose the condition (26) is satisfied. We

That is, the optimal ξk⋆ is inversely proportional to define the asymptotic SINR in (28) as a function g(τ ) of
the square of the long-term path loss of user k, which τ . Then the derivative of g(τ ) is derived as
compensates the long-term path loss in both the downlink
(1 − τ − α)ρ
WET phase and the uplink WIT phase. This optimal ξk⋆ g ′ (τ ) = −
enables users to asymptotically achieve a common rate (see (1 − ρ)K + (1 − τ − α)ρ
!
Theorem 5 later), thus ensuring fairness among users. M (M − K)αpdl βk2 ξk ρ
Substituting (36) into (26), we observe that M should log  −α)ρ 
.
σ 2 K + (1−τ
1−ρ
be much greater than some constant, given α and ρ. As
numerically shown in Section VII, for our analysis to hold, Clearly, the above derivative is negative, due to the assump-
⋆
the condition in (26) is well satisfied for M as small as tion of large M . Hence, we choose τZF → 0 to maximize
25. g(τ ).
2) Asymptotically Optimal Energy-Splitting: Here, we Suppose α = O(M −2ν ), for 0 < ν < 12 ; otherwise,
assume τ and α to be arbitrary. We first obtain the asymp- the asymptotically harvested energy in (27) decreases as
totically optimal ρ⋆ for Problem (P1 ) in the following M increases, which is suboptimal. Define ǫ(M ) such
lemma. that ǫ(M ) → 0 for sufficiently large M . From (28), the
Lemma 7. For ZF detection, when M satisfies (26), the asymptotic rate is rewritten as
!
asymptotically optimal ρ⋆ZF is given by
M 2(1−ν)
p β 2
ξ ρ
ekZF −−−→ 1−τ −ǫ(M ) log 1+  dl k k
√ R −ǫ(M))ρ 
⋆ K σ 2 K + (1−τ 1−ρ
ρZF (τ, α) = √ √ . (37) !
K + 1−τ −α M 2(1−ν) pdl βk2 ξk ρ
For MRC detection, when M satisfies (29), the asymp- −−−→ (1 − τ ) log 1 +  )ρ 
. (39)
σ 2 K + (1−τ
totically optimal energy-splitting fraction ρ⋆MRC that maxi- 1−ρ

mizes the achievable rate is arbitrary in (0, 1). We choose ν → 0 to maximize the asymptotic rate. Hence,
Proof: For ZF detection, given τ, α, the asymptotic α⋆ZF = O(M −2ν ), where ν → 0.
SINR in (28) is proportional to the function Lemma 9. For MRC detection, when M is sufficiently
ρ large to satisfy (29), the asymptotically optimal time al-
f (ρ) = (1−τ −α)ρ
. location for CE and WET is given by
K + 1−ρ


It can be shown that f (ρ) is a quasiconcave function of
⋆
τMRC −−−→ 0, α⋆MRC = O M −ϕ → 0, (40)
ρ. The optimal ρ⋆ that maximizes f (ρ) is obtained by where 0 < ϕ < 1 is arbitrary.
solving f ′ (ρ) = 0 where f ′ (ρ) is the derivative of f (ρ).
Proof: From (30), the asymptotically optimal τ and α
The solution is given by (37).
should be chosen as small as possible. To satisfy (29), we
For MRC detection, the asymptotic rate in (30) is
choose α to tend to zero at the rate of order of O (M −ϕ ),
independent of ρ. Hence, the energy-splitting coefficient
for 0 < ϕ < 1.
ρ can be arbitrary chosen in (0, 1).
For ZF detection, the optimal WET time α⋆ZF in
For ZF detection, we observe that the asymptotically
Lemma 8 approaches zero fairly slowly with the number
optimal ρ⋆ZF increases as the number of users K increases.
of antennas M , as will be verified numerically in Sec-
It implies that a higher fraction of harvested energy should
tion VII. For MRC detection, however, the optimal α⋆MRC
be used for CE for larger K. This is because the asymptotic
in Lemma 9 approaches zero faster with M , when ϕ is
rate is more interference-limited for large K, and hence
chosen to approach one.
more accurate CSI is required to decrease the interference
that comes from channel estimation error.
4) Asymptotically Maximal Minimum Rate: For both ZF
3) Asymptotically Optimal Time Allocation: The
detection and MRC detection, the asymptotically maximal
asymptotically optimal τ ⋆ and α⋆ is given in the
following Lemma 8 for ZF detection, and in Lemma 9 minimum achievable rate for Problem (P1 ) is thus given
in the following theorem.
for MRC detection.

Theorem 5. The asymptotically maximal minimum rate
for the proposed WET-MM system is
 !

 2
 M p
log 1 + σ2 (√K+1)2 PK 1 , for ZF
dl
Re −−−→ i=1 2
   βi

 M−1
log 1 + K−1 , for MRC.
(41)
Proof: For MRC, from Lemma 9 and Theorem 2, the
asymptotically maximal rate for user k is obtained in (41).
For ZF, from Lemma
√ 7 and Lemma 8, the asymptotically
K
optimal ρ⋆ZF = √K+1 . From Lemma 8 and Theorem 1, the
asymptotically maximal rate for user k is obtained in (41).
The fact that each user achieves the same asymptotically
maximal rate completes the proof.
Remark 3 (Fairness comparison). The proposed WET-MM
system asymptotically achieves the best possible rate fair-
ness among users (i.e., a common rate), while no fairness is
guaranteed in the OP-MM system. For the OP-MM system,
from (24), the rate achieved by users located far away may
be extremely low, since few energy is harvested.
Remark 4 (Asymptotic performance for large M and K).
Previously, the number of users K is fixed. Herein, we
let K scale up with M while keeping a constant ratio
K K
M , ζ ∈ (0, 1). As M, K → ∞ and M = ζ, similarly,
we obtain the asymptotic rate as follows
 ⋆  the achievable rates in (24) and (25) are functions of τ, α
e α pdl (1 − ζ) e and ρ. Hence, the asymptotically optimal τOT , α⋆OT and ρ⋆OT
R −−−→ log , R(ζ), (42)
c1 σ 2 ζ 2
1
PK 1
where c1 = K i=1 βi2 . We assume the path loss is
modeled as βk = β0 d−u k , where β0 is the loss for a distance
of one meter, and u > 1 is the path-loss exponent. We then
have the following claim on the quantity c1 .
Claim 1. Assuming the distance dk is independent and
uniformly distributed in the interval [a, b], where 0 < a <
b, it holds that c1 approaches a constant for large K, i.e.,
β −2 (b2u+1 −a2u+1 )
c1 → 0 (b−a)(2u+1) , as K → ∞.
Proof: From the path loss model, we have that c1 =
β0−2 PK 2u
i=1 Yi , where Yi = di . The probability density
K
function of random variable Yk is obtained as fY (y) =
y (0.5−u)/u 2u
2u(b−a) , for a ≤ y ≤ b2u , and fY (y) = 0 otherwise.
2u+1 2u+1
It is standard to show that E[Yi ] = (b(2u+1)(b−a)
−a )
. The
desired result is obtained by using the independence among
Yi ’s and the law of large numbers.
In the regime of large M and K, for a desired common
rate R0 , the asymptotically supporting number of users K
e
is given by ζ0 , where ζ0 is the solution to R(ζ) = R0 for
ζ. It can be shown that the positive solution ζ0 is unique.
In the ideal case, the H-AP has perfect CSI, without
requiring users to send pilots. Comparatively, the proposed
WET-MM system with larger K suffers inevitable perfor-
mance degradation, since users have to spend more time
10
and energy for sending pilots (as implied in Lemma 7).
However, the proposed WET-MM system achieves the
best achievable performance in the present optimization
framework with imperfect CSI, in which we vary the
allocation of time and energy resources, and the fraction
of harvested energy used for sending pilots.
B. Asymptotically Optimal Solutions for Ideal Case and
OP-MM System
1) Ideal Case: Similar to Section VI-A1, it can be
shown that the asymptotically optimal energy allocation
weights ξ ⋆ is given by (36). Thus, from (34) and (35), the
achievable rate is given by
ek,Ideal (α) −−−→
R
  



 (1−α) log 1+ αpdl M(M−K)  , for ZF
K
P
σ2 (1−α) 1
2 (43)

   βi
i=1

 M−1
(1−α) log 1+ K−1 , for MRC.
For ZF detection, similar to Lemma 8, it
can be shown
from (43), the optimal α⋆Ideal = O M −2ν → 0, where
ν > 0, and ν → 0. For MRC, similar to Lemma 9,
from (43), we have that the optimal α⋆Ideal = O (M −ν ) →
0, where 0 < ν < 1 is arbitrary.
2) OP-MM System: For both ZF and MRC detection,
⋆
that maximize the achievable rate can be obtained by a
numerical search.
VII. N UMERICAL R ESULTS
In this section, we present numerical results to validate
our results. We employ ZF detection given in this paper, as
well as MRC detection, see details in [33]. We set K = 2
and pdl = 1 Watt. The frame length is fixed as 5 µs, which
is normalized in the simulations. The carrier frequency is
5 GHz, and the bandwidth is 100 KHz. We set the power
spectrum density of noise as −170 dBm/Hz, which implies
the noise power σ 2 = −120 dBm. We use the long-term
fading model βi = 10−3 d−3 i , where the distance d1 = 6
m and d2 = 12 m. We use 1000 channel realizations in
the Monte Carlo simulation. The step size for τ , α and ρ
is chosen as 0.00025, 0.0005 and 0.0005, respectively.
First, we verify the asymptotically optimal solution for
ZF detection. We fix M = 200. By numerical search, the
optimal time for CE and for WET is τ ⋆ = 0.00825, α⋆ =
0.0760, respectively. From the analytic result (37) in
Lemma 7, the asymptotically optimal ρ⋆ = 0.5961.
From (41) in Theorem 5, the maximal minimum rate is thus
16.0096 bps/Hz. These results will be verified in Figs. 2,
Fig. 3 and Fig. 4.
For fixed ρ = 0.5965, Fig. 2 and Fig. 3 are the contour
plot of the rate against the uplink CE time τ and the
 11

1 16.5

0.9
2 16
0.8 4

0.7 15.5
6
2

Rate: R (bps/Hz)
0.6
WET Time: α

8 15
4
0.5 10

6
14.5
0.4 2 User 1: Simulation
12 8 User 2: Simulation
0.3 4
14 16.3491
14 User 1: Analytic LB
10 User 2: Analytic LB
0.2 Asymptotic LB
6
2 13.5
0.1
8

12 4
14 13
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Channel Estimation Time: τ Energy-Splitting Ratio: ρ

Fig. 2: Rate of user 1 versus CE time τ and WET time α. Fig. 4: Effect of energy-splitting fraction ρ on data rate.

1
in Table I. We use the normal notation for the analytically
0.9 asymptotic results, and use the notations with the bar head
2
0.8 4 for the numerical results. We observe that as M tends to
0.7
infinity, the optimal CE time τ̄ ⋆ tends to 0 quickly, while
6
2
the optimal WET time ᾱ⋆ tends to 0 at a slower rate. This
0.6
observation coincides with the results in Lemma 8. From
WET Time: α

8
4
0.5 10 Lemma 7 and Lemma 8, we obtain that the asymptotically
0.4 6
2
optimal energy-splitting ratio ρ⋆ = 0.5961 for M = 200,
12 8 4
which is approached by the numerically obtained optimal
0.3
15.9540
10 ρ̄⋆ in Table I.
14

0.2 Next, we simulate the achievable rates for ZF detection

6 2
0.1 and MRC detection. Fig. 5 compares the achievable rate
8

4
12

14
0.1 0.2
Fig. 3: Rate of user 2 versus CE time τ and WET time α.
downlink WET time α, for user 1 and 2, respectively.
Specifically, the maximal minimum rate is achieved as
15.9540 bps/Hz, at τ ⋆ = 0.00825, α⋆ = 0.0760. Further-
more, the achieved rate is 16.3491 bps/Hz and 15.9540
bps/Hz for user 1 and user 2, respectively, which is
almost the same. The asymptotically maximal minimum
rate is 16.0096 bps/Hz, which approximates the actually
maximal minimum rate 15.9540 bps/Hz closely. Hence, the
asymptotic result in Theorem 5 is numerically verified.
For fixed time allocation τ ⋆ = 0.00825, α⋆ = 0.0760,
the data rate is plotted in Fig. 4 against the energy-splitting
fraction ρ. We see that the rate is a quasi-concave function
of ρ, and the unique optimal energy-splitting fraction ρ⋆ ≈
0.6 is the star-marker point in Fig. 4. This coincides with
the analytic result in Lemma 7.
Moreover, we compare the optimal resource allocation
solutions for different M . The numerical results are shown
10
to that of the ideal case and of the OP-MM system, where
0.3 0.4 0.5 0.6 0.7 0.8 0.9
Channel Estimation Time: τ the x-axis is in the logarithm scale. For ZF detection, we
observe that the proposed WET-MM system achieves more
than 80% of the rate limit achieved by the ideal case
with perfect CSI. Also, we see that the proposed WET-
MM system achieves higher rates than the OP-MM sys-
tem. These observations numerically show the efficiency
of the proposed system. Moreover, we observe that the
analytically-obtained rate approaches that obtained from
simulation, even when M is as small as 10. The asymptotic
analysis is thus accurate even for the case of small M .
Finally, we see that the achievable rate for ZF detection is
higher than that for MRC detection, which is as expected.
Recall that the MM-DoRG κ is defined in (1) as the
asymptotic slope of rate R with respect to log M . We then
observe from Fig. 5 that the MM-DoRG of the proposed
WET-MM system achieves the same MM-DoRG as the
ideal case, i.e., κWET−MM = κIdeal = 2, which is double
of the MM-DoRG of the OP-MM system. Hence, the
proposed WET-MM system is numerically verified to be
optimal in terms of MM-DoRG.
Also, we observe from Fig. 5 that to achieve a max-
min rate at 10, 8, 6.4 bps/Hz for the maximal distance
 12

TABLE I: Optimal resource allocation versus M

M τ̄ ⋆ ᾱ⋆ ρ⋆ ρ̄⋆ easym.

R ⋆
Reanal.
⋆
R̄1 R̄2
25 0.03750 0.1455 0.6125 0.6425 10.4597 10.5162 12.1786 10.5635
50 0.02725 0.1180 0.6075 0.6250 12.3200 12.3327 13.4457 12.3319
100 0.01875 0.0945 0.6025 0.6025 14.1650 14.2022 14.7690 14.2114
200 0.00825 0.0760 0.5961 0.5965 16.0096 16.0098 16.3491 15.9540
400 0.00475 0.0580 0.5943 0.5950 17.8603 17.8677 18.0551 17.8744
600 0.00275 0.0515 0.5936 0.5930 18.9469 18.9490 19.0955 18.9520
800 0.00125 0.0505 0.5912 0.5915 19.7195 19.7147 19.8604 19.7240
1000 0.00075 0.0490 0.5901 0.5905 20.3196 20.3227 20.4010 20.3233

25 25
WET−MM (ZF): Lower Bound WET−MM (ZF): User 1
WET−MM (ZF): Asymp. LB WET−MM (MRC): User 1
WET−MM (ZF): Simulation WET−MM (ZF): User 2
20 20
WET−MM (MRC): Simulation WET−MM (MRC): User 2
Max-Min Rate: R (bps/Hz)

Ideal Case (ZF): Lower Bound Ideal Case (ZF): User 1

Rate: R (bps/Hz)
Ideal Case (ZF): Simulation Ideal Case (ZF): User 2
15 OP−MM (ZF): Simulation 15 OP−MM (ZF): User 1
OP−MM (ZF): User 2

10 10

5 5

0 0 1 2 3
0 0 1 2 3
10 10 10 10 10 10 10 10
Number of Antennas: M Number of Antennas: M

Fig. 5: Comparison of achievable rates versus M . Fig. 6: Comparison of rate fairness versus M .

d2 = 12 m, the OP-MM system requires about 400, is numerically shown to require fewer antennas at the H-
100, 25 antennas at the H-AP, while the proposed WET- AP, the number of which is roughly a square root of that
MM requires only 21, 10, 5 antennas, respectively, being required by the OP-MM system. The WET-MM system is a
roughly a square root of that required by OP-MM system. promising system to provide high data rate and overcome
Fig. 6 compares the user fairness. Using the derived the energy bottleneck for wireless devices. Nevertheless,
energy allocation weights for downlink WET in (36), the some issues related to practical implementation remain to
two users in the proposed WET-MM system asymptotically be addressed, such as the effect of antenna correlation,
achieve the same rate. In contrast, for the OP-MM system, the effect of imperfect channel reciprocity, other fading
the far user 2 has a much lower rate than the near user channel model with line-of-sight, etc.
1, even in the massive MIMO regime. Hence, better user
fairness is achieved by the proposed WET-MM system. A PPENDIX A
P ROOF OF L EMMA 1
VIII. C ONCLUSION PK
Let ξ ′ ∈ [0, 1], θi′ ∈ [0, 1], ∀k, i, subject to k=1 ξk′ +
This paper studies a WET-enabled massive MIMO sys- PM−Kk
i=1 θi = 1. Denote θ = [θ1 θ2 · · · θM−K ]T . Give the
tem with imperfect CSI. The minimum rate among users channel estimate G, b if the beamformer does not have the
is maximized by jointly optimizing the allocation of time structure in (8), we can write the beamformer as
and energy resources. The proposed WET-MM system
K q
X M−K
X p
achieves the maximum MM-DoRG achieved by the ideal b = bk
g b
w0 (G) ξk′ + θi ui (G), (44)
case with perfect CSI, which is double of the MM-DoRG kb
gk k 2
k=1 i=1
of the OP-MM system. Moreover, the WET-MM system
achieves the best possible fairness among users, since b M−K is an orthonormal basis for the
where {ui (G)} i=1
all users asymptotically achieve a common rate. Also, to orthogonal complement of the space spanned by {b gk } K
k=1 .
achieve a desired common rate for all users with a given b
From the property of MMSE estimation, gk is inde-
maximal AP-user distance, the proposed WET-MM system pendent of ek . Conditioned on gbk , gk is distributed as
 13

 
2
CN gbk , σe,k IM . Hence, we have former (44), we can always construct the beamformer (8) to
  asymptotically harvested more energy. Hence, the asymp-
H 2
EG|G bkH .
bk g
b gk gk = σe,k IM + g (45) totically optimal beamformer has the structure as in (8).
With the beamformer in (44), the expected harvested
energy by user k is rewritten from (7) as
Q′k (L, pCE ′ A PPENDIX B
k , α, ξk , θ)
h   i P ROOF OF L EMMA 2
= αpdl E b wH (G)E b b g g
k k
H
w 0 ( b
G)
G 0 G|G
" K !H
(a) Xq bk
g
Substituting (8) into (7), the harvested energy is
= αpdl EG b ξk′ 2
σe,k IM + g bkH ·
bk g " 
kbgk k 2 CE
k=1 Qk (L, pk , α, ξk ) = αEG E
!# " M−K !H b G| G
b
XK q Xp
′ bk
g b √
H √
ξk +αpdl EG b θi ui (G) · pdl ξk gkH gbk
2
pdl ξk gkH g bk X ξi gkH g
kb gk k 2 bi
k=1 i=1
!# 2
+
K q kb
gk k kb
gk k 2 kb
gi k 2

X b
g k
i6=k
2
σe,k IM + g bk
bk g H
ξk′ + αpdl · √ √
H
kbgk k 2 bk X ξi gkH g
pdl ξk gkH g bi
k=1 + +
" K !H kb
gk k 2 kbgi k 2
Xq b
g
i6=k
EGb ξk′
k 2
σe,k IM + g bkH ·
bk g p
X X pdl ξi ξj gkH g

H #
kbgk k 2 bi gkH g bj
k=1 . (49)
M−K
!# " M−K !H kb
gi k2 kb
gj k 2
Xp Xp i6=k j6=k
θi ui (G)b +αpdl E b b
θi ui (G) ·
G In the sequel, we investigate the four terms in (49).
i=1 i=1
M−K
!# Recall gbk = gk + ek and (45). Conditioned on the

X p b the harvested energy in the first term
2
σe,k bkH
bk g
IM + g b
θi ui (G) channel estimate G,
i=1
is rewritten as
  " 2
#  
(b) pdl ξk gkH g bk bkH Egk |bgk gk gkH g
pdl ξk g bk
= αpdl ξk′ 2
M βk − (M − 1)σe,k EG|G =
b 2 2
M−K
X kb
gk k kb
gk k
+ αpdl βk (1 − ξk′ ) + σe,k
2
θi (46) 2


= pdl ξk σe,k + g bkH gbk . (50)
i=1
 
where (a) is from (44) and (45), (b) is from (55) in bk ∼ CN 0M , (βk − σe,k 2
From the fact that g )IM , we
Appendix B, as well as the othogonality between ui (G)b  H   
bk = M βk − σe,k
bk g 2
bk . From the beamformer in (44), we construct a
and g have that Egbk g . Hence, the first
beamformer which has the structure as in (8) with weights term in (49) is obtained as
" " 2
##
ξj′ pdl ξk gkH g bk
ξj = PK . (47) αEG b EG|G b
′ gk k 2
kb
j=1 ξk  
2
With the beamformer in (8) with weights given by (47), = αpdl ξk M βk −(M −1)σe,k . (51)
the harvested energy is obtained in (55) in Appendix B. ei , kbggbiik2 , and ḡk , kb
Define g gk k 2 g bk . From (45), the
From (46), (47), (55) and (5), we have second term in (49) is rewritten as
Qk (L, pCE ′ CE ′  
H 
k ,α, ξk ) − Qk (L, pk , α, ξk , θ) √
X √ξi gH g
pdl ξk gkH g bk b
k i 
M−K
X αEG 
b EG|G

αpdl ξk′ (M − 1)βk LpCE
k θi
b
kb
gk k 2 kb
gi k 2
= βk i6=k
P
K
p Xp  H 
i=1 (σ 2 + βk LpCEk ) ξj′ 2
ek gei
j=1
= αpdl ξk σe,k ξi Egek ,egi g
M−K
X   i6=k
σ2 p Xp   (a)
+ βk θi αpdl − 2 + αpdl ξk ξi Eḡk ,egi ḡkH g
ei = 0, (52)
i=1
σ + βk LpCE
k
i6=k
≥ 0, as M → ∞, (48)
ek and ḡk are independent
where (a) is from the fact that g
where the inequality is from the fact that the M -dependent zero-mean random vectors, for any i 6= k.
term dominates the minus term. That is, for any beam- The third term in (49), which is the conjugate of the
 14

second term in (49), is similarly obtained as It can be shown that the random variable Zk ,
 

2
bkH g
g bkfollows central chi-square distribution with
√ √ Hb H
2
βk −σe,k
p ξ g Hb X
g ξi g g i
 
αEG 
b EG|G b
 dl k k k k  = 0. 2M degrees of freedom. Then, we have E Z1k =
kbgk k 2 kb
gi k 2
i6=k 1
2(M−1) . Hence, it holds that
(53)    
1 2 1
The fourth term in (49) is rewritten as E = E
2 )
gkH g
pk |b bk | pk (βk − σe,k Zk
  p
H  1
X X pdl ξi ξj gkH g bi gkH g bj = 2 ). (58)
αEG 
b EG|G b
  pk (M − 1)(βk − σe,k
kb
gi k2 kb
gj k 2
i6=k j6=k Substituting (57) and (58) into (56), we obtain that
(a) X H 2
= αpdl EG b ξi gei (σe,k IM + g bkH )e
bk g gi eMRC = (1 − τ − α)
Rk
i6=k  
X X p  2


+ ξi ξj g 2
eiH (σe,k IM + bkH )e
bk g
g gj  pk (τ, α, ρ)(M − 1) βk − σe,k (α, ρ) 
log 1+ P 2 .
i6=k j6=i,k pi (τ, α, ρ)βi + pk (τ, α, ρ)σe,k (α, ρ) + σ2
(b) X i6=k
= αpdl βk ξi , (54) (59)
i6=k
The result in (21) is obtained, by substituting the error
where (a)
is from (45), and (b) is from the fact that variance in (18) into (59).
bkH = (βk − σe,k
bk g 2 ei and g ej are inde-
EGb g )IM , and g
pendent zero-mean random vectors, for any i 6= j.
Substituting (51), (52), (53) and (54) into (49), we obtain A PPENDIX D
the harvested energy as P ROOF FOR L EMMA 6
 2

Qk (L, pCE
k , α, ξk ) = αpdl ξk M βk −(M −1)σe,k We rewrite the asymptotic achievable rate in (28) for ZF
+ αpdl βk (1 − ξk ). (55) detection as follows,


RekZF −−−→ (1 − τ − α) log 1 + Ck (τ, α, ρ)ξk , (60)
Substituting (5) into (55), the harvested energy is in (16).
where Ck (τ, α, ρ) is the multiplicative coefficient of ξk in
A PPENDIX C the logarithm of (28).
P ROOF OF L EMMA 5 We consider the case in which the first (K − 1) users
H
b
gk b
gi eZF and the K-th user achieves a higher
With MRC, ak = g gi , kb
bk . Define e gk k . The achievable have the same rate R
rate of user k is written as e ZF
rate, i.e., R1 = R eZF = · · · = R e ZF = R eZF < R e ZF .
2 K−1 K
e MRC Clearly, one can always increase the minimum rate eZF
R
Rk = (1 − τ − α)
   K −1  by increasing ξ1 , ξP 2 , · · · , ξK−1 and decreasing ξK , subject
 P PK
 K
 
 g i |2 +
pi |e 2
pi σe,i + σ2  
to the constraint i=1 ξi = 1. Then the minimum rate
      eZF = R eZF .
 i=1,i6=k i=1   among users is maximized when R
log 
1 +  E  .
K

  
 p k |b
g H b
g k | 
   The same argument can be extended to other cases where


k 
 less than (K − 1) users have the same UL rate. That is,
the minimum rate is maximized when all users achieves
(56)
the same rate; otherwise, one can always increase the
It can be easily shown that gei ∼ CN (0, βi − σe,i 2
). It minimum rate by adjusting ξk ’s. Let R e ZF = (1 − τ −
is further noted that the conditional probability density α) log(1+γ), where P the common SINR γ = Ck (τ, α, ρ)ξk .
K
function (pdf) f (e gi |bgk ) = f (e gi ), where f (e gi ) is the From the constraint i=1 ξi = 1, we obtain that
marginal pdf of e gi . Then we have γ
 K  ξk⋆ =
 P PK
 Ck (τ, α, ρ)

 g i |2 +
pi |e pi σe,i2
+ σ2  
  1 1
i=1,i6=k i=1 = PK = PK 1 .
E = C (τ, α, ρ) 1
β 2

 pk |b gkH g
bk | 

k i=1 Ci (τ,α,ρ) k i=1 βi2

 
 ⋆
Hence, the asymptotically optimal ξk only depends on the
 
XK   long-term path loss βi ’s of all users.
 2 2 1
pi βi + pk σe,k + σ E . (57) For MRC detection, by using the Karush-Kuhn-Tucker
i=1,i6=k
gkH g
pk |b bk | conditions [34], it can be shown that the minimum rate
 15

is maximized when all users achieves the same rate. The [23] X. Kang, C. K. Ho, and S. Sun, “Full-duplex wireless pow-
energy allocation weight ξk⋆ is thus given by (36). ered communication network with energy causality,” to ap-
pear in IEEE Trans. Wireless Commun. Available on-line at
arxiv.org/abs/1404.0471.
R EFERENCES [24] F. Rusek, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors, and
F. Tufvesson, “Scaling up MIMO: opportunities and challenges with
[1] N. Shinohara, “Power without wires,” IEEE Microw. Mag., vol. 12, very large arrays,” IEEE Signal Process. Mag., pp. 40–60, January
pp. S64–S73, December 2011. 2013.
[2] S. Bi, C. K. Ho, and R. Zhang, “Wireless powered communication: [25] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang,
opportunities and challenges,” to appear in IEEE Commun. Mag. “An overview of massive MIMO: benefits and challenges,” IEEE
[Online]. Available: http://arxiv.org/abs/1408.2335. Journal of Selected Topics in Signal Processing, vol. 8, pp. 742–
[3] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous 758, October 2014.
wireless information and power transfer,” IEEE Trans. Wireless [26] X. Chen, C. Yuen, and Z. Zhang, “Wireless energy and information
Commun., vol. 12, pp. 1989–2001, May 2013. transfer tradeoff for limited feedback multi-antenna systems with
[4] L. R. Varshney, “Transporting information and energy simultane- energy beamforming,” IEEE Trans. Veh. Technol., vol. 63, pp. 407–
ously,” in IEEE Int. Symp. Inf. Th., pp. 1612–1616, 2008. 412, January 2014.
[5] X. Zhou, R. Zhang, and C. K. Ho, “Wireless information and power [27] X. M. Chen, X. M. Wang, and X. F. Chen, “Energy efficient
transfer: architecture design and rate-energy tradeoff,” IEEE Trans. optimization for wireless information and power transfer in large
Commun., vol. 61, pp. 4757–4767, November 2013. MIMO employing energy beamforming,” IEEE Wireless Commun.
[6] L. Liu, R. Zhang, and K. C. Chua, “Wireless information transfer Lett., vol. 2, pp. 667–670, December 2013.
with opportunistic energy harvesting,” IEEE Trans. Wireless Com- [28] L. Liu, R. Zhang, and K. C. Chua, “Multi-antenna wireless powered
mun., vol. 12, pp. 288–300, Jan. 2013. communication with energy beamforming,” IEEE Trans. Commun.,
[7] P. Grover and A. Sahai, “Shannon meets Tesla: wireless information vol. 62, pp. 4349–4361, December 2014.
and power transfer,” in IEEE Int. Symp. Inf. Th., pp. 2363–2367, [29] G. Yang, C. K. Ho, and Y. L. Guan, “Dynamic resource allocation
2010. for multiple-antenna wireless power transfer,” IEEE Trans. Signal
[8] K. Huang and E. G. Larsson, “Simultaneous information and Process., vol. 62, pp. 3565–3577, July 2014.
power transfer for broadband wireless systems,” IEEE Trans. Signal
[30] J. Xu and R. Zhang, “Energy beamforming with one-bit feedback,”
Process., vol. 61, pp. 5972–5986, December 2013. IEEE Trans. Signal Process., vol. 62, pp. 5370–5381, October 2014.
[9] D. W. K. Ng, E. S. Lo, and R. Scholar, “Wireless information and [31] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral
power transfer: energy efficiency optimization in OFDMA systems,” efficiency of very large multiuser MIMO systems,” IEEE Trans.
IEEE Trans. Wireless Commun., vol. 12, pp. 6352–6370, December Commun., vol. 61, pp. 1436–1449, April 2013.
2013. [32] S. M. Kay, Fundamentals Of Statistical Signal Processing: Estima-
[10] X. Zhou, R. Zhang, and C. K. Ho, “Wireless information and tion Theory. NJ, USA: Prentice Hall, 1993.
power transfer in multiuser OFDM systems,” IEEE Trans. Wireless [33] G. Yang, C. K. Ho, R. Zhang, and Y. L. Guan, “Throughput
Commun., vol. 13, pp. 2282–2294, April 2014. optimization for massive MIMO systems powered by wireless
[11] J. Xu, L. Liu, and R. Zhang, “Multiuser MISO beamforming for energy transfer,” Available on-line at arxiv.org/abs/1403.3991.
simultaneous wireless information and power transfer,” IEEE Trans.
[34] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
Signal Process., vol. 62, pp. 4798–4810, September 2014. University Press, 2004.
[12] H. Ju and R. Zhang, “A novel model switching scheme utilizing
random beamforming for opportunistic energy harvesting,” IEEE
Trans. Wireless Commun., vol. 13, pp. 2150–2162, April 2014.
[13] A. M. Fouladgar and O. Simeone, “On the transfer of information
and energy in multi-user systems,” IEEE Commun. Lett., vol. 16,
pp. 1733–1736, November 2012.
[14] A. A. Nasir, X. Zhou, S. Durrani, and R. A. Kennedy, “Relay pro-
tocals for wireless energy harvesting and information processing,”
IEEE Trans. Wireless Commun., vol. 12, pp. 3622–3636, July 2013.
[15] C. Shen, W.-C. Li, and T.-H. Chang, “Simultaneous information
and energy transfer: a two-user MISO interference channel case,”
in Proc. IEEE Global Commun. Conf. (GLOBECOM), pp. 3862–
3867, 2012.
[16] J. Park and B. Clerckx, “Joint wireless information and energy
transfer in a two-user MIMO interference channel,” IEEE Trans.
Wireless Commun., vol. 12, pp. 4210–4221, August 2013.
[17] K. Huang and V. K. N. Lau, “Enabling wireless power transfer in
cellular networks: architecture, modeling and deployment,” IEEE
Trans. Wireless Commun., vol. 13, pp. 902–912, February 2014.
[18] S. Lee, R. Zhang, and K. Huang, “Opportunistic wireless energy
harvesting in cognitive radio networks,” IEEE Trans. Wireless
Commun., vol. 12, pp. 4788–4799, September 2013.
[19] H. Ju and R. Zhang, “Throughput maximization in wireless powered
communication networks,” IEEE Trans. Wireless Commun., vol. 13,
pp. 418–428, January 2014.
[20] Product Datasheet, P2110-915MHz RF Powerharvester
Receiver, Powercast Corporation. Available on-line at
www.powercastco.com/PDF/P2110-datasheet.pdf.
[21] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan, and M. C.
Reed, “Femtocells: past, present, and future,” IEEE J. Sel. Areas
Commun., vol. 30, pp. 497–508, April 2012.
[22] Product Datasheet, 2.4 GHz Ultra-Low Power Radio, Interuniversity
Microelectronics Centre (IMEC). Available on-line at
www2.imec.be/be en/research/wireless-communication/ultralow-power-wireless-communic.html.